pkg/R/F_Neighbors2.R

   1 #' @include Forecaster.R
   2 #'
   3 #' Neighbors2 Forecaster
   4 #'
   5 #' Predict tomorrow as a weighted combination of "futures of the past" days.
   6 #' Inherits \code{\link{Forecaster}}
   7 #'
   8 Neighbors2Forecaster = R6::R6Class("Neighbors2Forecaster",
   9     inherit = Forecaster,
  10
  11     public = list(
  12 #       predictSerie = function(data, today, memory, horizon, ...)
  13 #       {
  14 #           # Parameters (potentially) computed during shape prediction stage
  15 #           predicted_shape = self$predictShape(data, today, memory, horizon, ...)
  16 ##          predicted_delta = private$.pjump(data,today,memory,horizon,private$.params,...)
  17 #           # Predicted shape is aligned it on the end of current day + jump
  18 ##          predicted_shape+tail(data$getSerie(today),1)-predicted_shape[1]+predicted_delta
  19 #           predicted_shape
  20 #       },
  21         predictShape = function(data, today, memory, horizon, ...)
  22         {
  23             # (re)initialize computed parameters
  24             private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
  25
  26             # Do not forecast on days with NAs (TODO: softer condition...)
  27             if (any(is.na(data$getCenteredSerie(today))))
  28                 return (NA)
  29
  30             # Determine indices of no-NAs days followed by no-NAs tomorrows
  31             fdays = getNoNA2(data, max(today-memory,1), today-1)
  32
  33             # Get optional args
  34             simtype = ifelse(hasArg("simtype"), list(...)$simtype, "mix") #or "endo", or "exo"
  35             kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
  36             if (hasArg(h_window))
  37             {
  38                 return ( private$.predictShapeAux(data,
  39                     fdays, today, horizon, list(...)$h_window, kernel, simtype, TRUE) )
  40             }
  41
  42             # Indices of similar days for cross-validation; TODO: 45 = magic number
  43             sdays = getSimilarDaysIndices(today, limit=45, same_season=FALSE)
  44
  45             cv_days = intersect(fdays,sdays)
  46             # Limit to 20 most recent matching days (TODO: 20 == magic number)
  47             cv_days = sort(cv_days,decreasing=TRUE)[1:min(20,length(cv_days))]
  48
  49             # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
  50             errorOnLastNdays = function(h, kernel, simtype)
  51             {
  52                 error = 0
  53                 nb_jours = 0
  54                 for (i in seq_along(cv_days))
  55                 {
  56                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  57                     prediction = private$.predictShapeAux(data,
  58                         fdays, cv_days[i], horizon, h, kernel, simtype, FALSE)
  59                     if (!is.na(prediction[1]))
  60                     {
  61                         nb_jours = nb_jours + 1
  62                         error = error +
  63                             mean((data$getSerie(cv_days[i]+1)[1:horizon] - prediction)^2)
  64                     }
  65                 }
  66                 return (error / nb_jours)
  67             }
  68
  69             if (simtype != "endo")
  70             {
  71                 h_best_exo = optimize(
  72                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="exo")$minimum
  73             }
  74             if (simtype != "exo")
  75             {
  76                 h_best_endo = optimize(
  77                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="endo")$minimum
  78             }
  79
  80             if (simtype == "endo")
  81             {
  82                 return (private$.predictShapeAux(data,
  83                     fdays, today, horizon, h_best_endo, kernel, "endo", TRUE))
  84             }
  85             if (simtype == "exo")
  86             {
  87                 return (private$.predictShapeAux(data,
  88                     fdays, today, horizon, h_best_exo, kernel, "exo", TRUE))
  89             }
  90             if (simtype == "mix")
  91             {
  92                 h_best_mix = c(h_best_endo,h_best_exo)
  93                 return(private$.predictShapeAux(data,
  94                     fdays, today, horizon, h_best_mix, kernel, "mix", TRUE))
  95             }
  96         }
  97     ),
  98     private = list(
  99         # Precondition: "today" is full (no NAs)
 100         .predictShapeAux = function(data, fdays, today, horizon, h, kernel, simtype, final_call)
 101         {
 102             fdays = fdays[ fdays < today ]
 103             # TODO: 3 = magic number
 104             if (length(fdays) < 1)
 105                 return (NA)
 106
 107             # Neighbors: days in "same season"
 108             sdays = getSimilarDaysIndices(today, limit=45, same_season=TRUE, data)
 109             indices = intersect(fdays,sdays)
 110             levelToday = data$getLevel(today)
 111             distances = sapply(indices, function(i) abs(data$getLevel(i)-levelToday))
 112             same_pollution = (distances <= 2)
 113             if (sum(same_pollution) < 1) #TODO: 3 == magic number
 114             {
 115                 same_pollution = (distances <= 5)
 116                 if (sum(same_pollution) < 1)
 117                     return (NA)
 118             }
 119             indices = indices[same_pollution]
 120
 121             #TODO: we shouldn't need that block
 122             if (length(indices) == 1)
 123             {
 124                 if (final_call)
 125                 {
 126                     private$.params$weights <- 1
 127                     private$.params$indices <- indices
 128                     private$.params$window <- 1
 129                 }
 130                 return ( data$getSerie(indices[1])[1:horizon] ) #what else?!
 131             }
 132
 133             if (simtype != "exo")
 134             {
 135                 h_endo = ifelse(simtype=="mix", h[1], h)
 136
 137                 # Distances from last observed day to days in the past
 138                 serieToday = data$getSerie(today)
 139                 distances2 = sapply(indices, function(i) {
 140                     delta = serieToday - data$getSerie(i)
 141                     mean(delta^2)
 142                 })
 143
 144                 sd_dist = sd(distances2)
 145                 if (sd_dist < .Machine$double.eps)
 146                 {
 147 #                   warning("All computed distances are very close: stdev too small")
 148                     sd_dist = 1 #mostly for tests... FIXME:
 149                 }
 150                 simils_endo =
 151                     if (kernel=="Gauss")
 152                         exp(-distances2/(sd_dist*h_endo^2))
 153                     else
 154                     {
 155                         # Epanechnikov
 156                         u = 1 - distances2/(sd_dist*h_endo^2)
 157                         u[abs(u)>1] = 0.
 158                         u
 159                     }
 160             }
 161
 162             if (simtype != "endo")
 163             {
 164                 h_exo = ifelse(simtype=="mix", h[2], h)
 165
 166                 M = matrix( nrow=1+length(indices), ncol=1+length(data$getExo(today)) )
 167                 M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
 168                 for (i in seq_along(indices))
 169                     M[i+1,] = c( data$getLevel(indices[i]), as.double(data$getExo(indices[i])) )
 170
 171                 sigma = cov(M) #NOTE: robust covariance is way too slow
 172 #               sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
 173                 sigma_inv = MASS::ginv(sigma)
 174
 175                 # Distances from last observed day to days in the past
 176                 distances2 = sapply(seq_along(indices), function(i) {
 177                     delta = M[1,] - M[i+1,]
 178                     delta %*% sigma_inv %*% delta
 179                 })
 180
 181                 sd_dist = sd(distances2)
 182                 if (sd_dist < .25 * sqrt(.Machine$double.eps))
 183                 {
 184 #                   warning("All computed distances are very close: stdev too small")
 185                     sd_dist = 1 #mostly for tests... FIXME:
 186                 }
 187                 simils_exo =
 188                     if (kernel=="Gauss")
 189                         exp(-distances2/(sd_dist*h_exo^2))
 190                     else
 191                     {
 192                         # Epanechnikov
 193                         u = 1 - distances2/(sd_dist*h_exo^2)
 194                         u[abs(u)>1] = 0.
 195                         u
 196                     }
 197             }
 198
 199             similarities =
 200                 if (simtype == "exo")
 201                     simils_exo
 202                 else if (simtype == "endo")
 203                     simils_endo
 204                 else #mix
 205                     simils_endo * simils_exo
 206
 207             prediction = rep(0, horizon)
 208             for (i in seq_along(indices))
 209                 prediction = prediction + similarities[i] * data$getSerie(indices[i]+1)[1:horizon]
 210             prediction = prediction / sum(similarities, na.rm=TRUE)
 211
 212             if (final_call)
 213             {
 214                 private$.params$weights <- similarities
 215                 private$.params$indices <- indices
 216                 private$.params$window <-
 217                     if (simtype=="endo")
 218                         h_endo
 219                     else if (simtype=="exo")
 220                         h_exo
 221                     else #mix
 222                         c(h_endo,h_exo)
 223             }
 224
 225             return (prediction)
 226         }
 227     )
 228 )